search

chuanqi305 / MobileNet-SSD

Caffe implementation of Google MobileNet SSD detection network, with pretrained weights on VOC0712 and mAP=0.727.
MIT License
2k stars 1.18k forks source link

Accuracy too low for custom dataset #191

Open funmonty opened 3 years ago

funmonty commented 3 years ago

Hi,

I have done transfer learning for object detection using the pretrained model given in the website. These are the parameters for my scenario :

Num of classes : 2 Dataset : Custom Num of train images : 6700 Num of test images : 1500 Iterations : 5000 max_lr : 0.001

With these parameters, I am able to achieve only 0.1732 accuracy. Can you please point out why it is too low and where should I tweak it? How many iterations would be ideal?